AMD Computex Presentation

[BoFox]

Here's what 2 RX 480s in Crossfire can do, Ryan notes "Are we still going to have vendors seriously pitch multi-GPU given all of the problems in the past year?"

Now we need to do "forensics" on this..  analyze the ordinary Radeon scaling in this game.  Do the recent Radeon cards generally scale at 100% or just 50% or what, for this game? 

From this, we can roughly estimate the true speed of a single RX 480. 

Extrapolating from this:

http://www.tweaktown.com/articles/7569/d...ndex3.html

We can see that 2x Fury X cards perform 61.5% faster than a single Fury X.

Let's just round it up to 62% scaling, assuming the the scaling has improved slightly (although air-cooled cards would be more temperature-restricted than water-cooled Fury X cards). 

62.5 fps in Crossfire, divided by 1.62 = 38.6 fps for a single RX 480, roughly!

Now, GTX 1080's 58.7 fps being 52% faster than a single RX 480's extrapolated 38.6 fps gives us some hint ( 0.657 of GTX 1080 ),
if we can find the most AMD-preferred AoS benchmark with settings that make the Radeon shine the most, like what AMD would try to use.  Tomshardware's bench will just be used for now, from http://www.tomshardware.com/reviews/nvid...585-3.html    NAH, SCRATCH THAT - I'm using this one from Hardware.fr, which shows the Fury X and 290X as even closer to that of 1080, which is probably closer to what AMD used:
http://www.hardware.fr/articles/948-16/b...arity.html

The factory-overclocked 290X being roughly the same as the 390X yields 46 fps, whereas the 1080 yields 64 fps.  There, the 1080 is 39.1% faster than the pre-oc'ed 290X.
( 0.71875 of GTX 1080 )

Now, we could probably safely assume that the RX 480 falls short of the 390X by quite a bit.  If the same benchmark by Hardware.fr with the exact same settings was used by AMD for the slideshow (while RX 480's scaling is roughly the same as that of the Fury X), then get ready for this:

The factory-oc'ed 290X (much like a 390X) would be 9% faster than the RX 480. 



On the other hand, if I use Tomshardware's benchmark instead of Hardware.fr's..:
GTX 1080:  66.1 fps
R9 390X:  38.6 fps  ( 0.584 of GTX 1080 )

If TomHardware's bench and settings were used by AMD, it would turn the tides over, inferring that RX 480 is 12.6% faster than the 390X.  AMD in all of their engineering wisdom would prefer benchmark settings more like the one that Hardware.fr used above for obvious reasons.

However, this could still imply that Polaris does indeed have some further Async efficiency benefits (additional DX12 optimizations and refinements), but I wouldn't expect so much there.  If this is the case, then it would just mean that for most other games, Polaris would not have such an edge over the 390X.  Nevertheless we have to keep Hardware.fr's charts in mind (while Hardware.fr generally painted the Radeon cards in better light than most other hardware review sites over the past few years when it came to the overall, aggregate averages against the Geforce rivals). 

The generic TFLOPs figure places the RX 480 in between the 390 and the 390X, yet there seems to be something else holding it back - perhaps the compressed bandwidth of only 256GB/s ("Amazing" 20% compression efficiency making it more like 307GB/s, compared to 390X's which is still yet 25% higher than 307GB/s) penalizes the card a bit.  Using a generic rule of thumb of at least 2% performance hit for every 10% reduction in bandwidth, which is a bit modest and conservative with the front-end/back-end to bandwidth ratio of flagship GPUs of today which is more like 3% rather than 2%, I'd say that this 20% reduction in effective bandwidth usage translates to at least around 5% effective loss of performance compared to using 390X's identical memory uarch.  

If it's still slower than the 390, rather than just the 390X, then there could be something else holding it back, like the ROPs being cut down to pathetic 32, from 64 (which I highly doubt AMD would do)....   The number of TMUs could've been cut down even more, proportion-wise (which seems to be the norm these days, as the shader count is the most important thing).  
https://www.techpowerup.com/gpudb/2848/radeon-rx-480
144 TMUs rumored, hmmm...   182.3 GTexels/s, compared to R9 390X's 184.8 GTexel/s - no change there at all, really.  So I'm puzzled..

We'll see soon anyway - too much speculation and hot air that I need a fan..   My bet is that it'd still need considerable overclocking to truly beat a 390X overall.